Diagnostic Sufficiency as Identifiability: Predicting Which Marker Panels Will Work from a Kinetic Model

A blood marker is a concentration, and a concentration is the net result of several rates: how fast the tissue makes or sheds the molecule, how fast the body clears it, and, in some diseases, how fast the diseased tissue traps it. A single marker fails as a diagnostic when the disease-driven rate cannot be separated from the rates that simply differ from person to person. Amyloid-beta 42 in Alzheimer's disease is the standard example. Plasma levels fall as the brain traps the peptide in plaques, but people also differ in how much they produce, so absolute plasma Aβ42 is weakly informative. Dividing by Aβ40, a sister peptide made by the same machinery that does not aggregate, cancels the production term. This rescue is reported in Highly sensitive blood-based biomarkers detection of beta-amyloid and phosphorylated-tau181 for Alzheimer's disease, where the ratio reaches AUC 0.8343. The ratio was found empirically, though, and so were nearly all the multi-marker panels in this literature.

This note proposes making that reasoning explicit and predictive. Write an ordinary differential equation (ODE) model, meaning a set of rate equations for how each molecule's concentration changes over time, for the system that produces the markers. Put a population distribution on the "nuisance" rates that vary between healthy people. Then ask, before collecting any labels, whether the disease parameter is identifiable from a candidate panel. Identifiable here means it can be recovered from the measurements once the nuisance variation is accounted for. The central claim is that single-marker insufficiency is the same thing as practical non-identifiability of the disease parameter. If that holds, a quantity computed from the mechanistic model should predict, panel by panel, how well a classifier trained on real labeled data will do. It should also say which small panel to measure.

Core Claim

Let $\theta$ be the disease-relevant rate, for example a sequestration constant or the strength of a constitutive pathway bypass. Let $\eta$ be the vector of nuisance rates (production, clearance, expression levels) that vary across individuals. Let $I$ be the Fisher information matrix of a candidate panel's measurements with respect to $(\theta, \eta)$, computed from the ODE's sensitivities. The information the panel carries about the disease after the nuisance has been accounted for is the Schur complement

$$I_{\text{eff}}(\theta) = I_{\theta\theta} - I_{\theta\eta}\, I_{\eta\eta}^{-1}\, I_{\eta\theta}.$$

This is the standard "efficient information" from semiparametric statistics. A marker can respond strongly to disease, with large $I_{\theta\theta}$, and still have $I_{\text{eff}} \approx 0$ when its disease response points in the same direction as a nuisance response. That is the overlap zone described for CA19-9 by Arakawa et al. and for CA125 in Guan et al.. A ratio or a panel works when it adds a measurement that cancels the nuisance direction.

Because real nuisance distributions are broad and the models are nonlinear, the local Fisher quantity is only a screening tool. The main quantity is a simulation-based "Bayes AUC" for each panel. To get it, simulate healthy and diseased individuals from the ODE with population priors, train an amortized posterior estimator (a neural network that learns once to map observations to a distribution over parameters, so later inference costs one forward pass), and score how well the posterior over $\theta$ separates the two groups. The falsifiable prediction is that the ranking of panels by this simulated score matches the ranking by held-out empirical AUC on real data.

Relation to Prior Work

Three ODE papers motivate the claim by showing single-marker failure as a kinetic phenomenon in cell lines. Decoupling of the PI3K Pathway via Mutation Necessitates Combinatorial Treatment in HER2+ Breast Cancer builds an ODE model of HER2–PI3K–AKT–S6RP signaling. It shows that PIK3CA mutations decouple p-S6RP from receptor control, so p-S6RP under lapatinib misreads mutant cells. Model-based analysis of response and resistance factors of cetuximab treatment in gastric cancer cell lines fits 230 kinetic parameters, finds MET-driven bypass, and uses profile likelihoods to show that many parameters are practically non-identifiable. Pharmacodynamic model of PARP1 inhibition and global sensitivity analyses can lead to cancer biomarker discovery is the closest in spirit. It ranks model parameters by global sensitivity and nominates them as biomarker candidates.

This proposal differs from all three in one specific way. They ask which parameters matter for the outcome, or which parameters can be estimated. We ask which measurable panels make the outcome-relevant parameter recoverable despite population nuisance variation, and we test that prediction against real labeled classification performance. Sensitivity alone is the wrong criterion. A highly sensitive marker that shares its sensitivity direction with a nuisance rate is exactly the marker that fails.

The inference machinery comes from An amortized approach to non-linear mixed-effects modeling based on neural posterior estimation. That work trains conditional normalizing flows on ODE simulations to infer individual and population parameters, and reports 97% interval coverage against 72% for the standard first-order estimation method on a 47-patient sunitinib dataset. It also warns that prior misspecification biases results, which constrains our design. We reuse its estimator but point it at a different target: a disease parameter, with the nuisance marginalized out, and comparisons across candidate panels. Pharmacokinetic fitting is not the goal.

The empirical panel literature supplies the benchmark and the foil. It includes the ILD random forest in Identification of routine blood derived hematological and lipid indices in ILD through machine learning (test AUC 0.885, external 0.849, NHR alone 0.793), the gastric peritoneal dissemination nomogram in Clinical Significance of Serum CA125, CA19-9, CA72-4, and Fibrinogen-to-Lymphocyte Ratio (AUC 0.910), and recursive feature elimination and LASSO panels in the ovarian LC/MS study and the LUSC metabolomics study. These studies found good panels by searching labeled data. None predicts panel quality before labels exist, and the external-validation drop in the ILD study (0.885 to 0.849) is the kind of shift a mechanism-grounded criterion should resist.

Research Program

The work has three parts.

The first part builds the evaluator. We extend the published HER2 signaling ODE, implemented in Catalyst (Catalyst), with an explicit split between the disease parameter $\theta$ and the nuisances $\eta$. Here $\theta$ is the constitutive PI3K activity contributed by the mutation. The nuisances are line-specific HER2, HER3, PI3K, AKT, and S6 abundances and turnover rates. We fit nuisance priors to the RPPA data from wild-type lines only, so no mutation labels touch the priors. For each candidate panel (a set of measured species and time points), we compute $I_{\text{eff}}$ and the simulated Bayes AUC. Candidates run from single proteins up to three-protein, two-time-point panels drawn from the species the model represents.

The second part tests the claim on real labels. The HER2+ RPPA panel (Decoupling of the PI3K Pathway) has 179 proteins, eight time points from 30 minutes to 72 hours, and 15 lines with known PIK3CA status. On each candidate panel we train a simple classifier for mutation status and for combination-benefit status, which separates the 5/5 and 6/6 responding mutant lines from the 1/17 responding wild-type lines. We score it by leave-one-line-out AUC. The headline statistic is the Spearman correlation across panels between simulated Bayes AUC and empirical held-out AUC. A second test uses the cetuximab immunoblot panel (Model-based analysis of response and resistance factors), where the disease parameter is MET bypass. We ask whether the criterion predicts which of pEGFR, pERK, and pAKT, and at which times, discriminates resistant from sensitive dynamics. Only two lines are available there, so this is a qualitative replication rather than a statistical test.

The third part is a clinical-scale check that runs on a reduced model and published numbers before any new data are needed. A two-compartment brain–plasma amyloid model, with shared production, aggregation that consumes only Aβ42, efflux, and clearance, should reproduce the reported ordering without fitting to labels: absolute Aβ42 is weak, Aβ42/Aβ40 is substantially better, and a marker on an independent axis such as p-tau181 adds information. If we can obtain the 116-participant digital ELISA cohort, the same panel-ranking test runs there. If we cannot, the clinical arm stays a qualitative consistency check and we say so.

Baselines

The naive reference is the best single marker, chosen on training lines and scored leave-one-line-out on the RPPA data. For the HER2 question that is p-S6RP under lapatinib, which the source paper already identifies as insufficient. In the amyloid check it is absolute Aβ42. Any proposed panel must clearly beat this, or the question is moot.

The strongest existing method is data-driven panel search followed by a flexible classifier. That means LASSO and recursive feature elimination selection feeding a random forest, which is the recipe behind the ILD model (test AUC 0.885 in the ILD study) and the LUSC signature (validation AUC 0.944 in the LUSC study). For the HER2 RPPA mutation-classification task itself, no reported result was found on Lacuna, so we will establish it ourselves. The search runs nested inside each leave-one-line-out fold, over the same candidate panel sizes the mechanistic criterion is allowed, with the same hyperparameter budget as the posterior estimator's training: equal CPU-hours, and an equal number of configurations tried. The proposal beats it if, at fixed panel size of two or three markers, the panel chosen by the mechanistic criterion without labels reaches held-out AUC at least as high as the panel chosen by label-driven search. It must also lose less AUC when evaluated on lines held out by a different grouping, namely HER2 amplification level, which plays the role of the cohort shift seen in the ILD external validation.

A third baseline is log-contrast logistic regression on log-concentrations with L1 penalty. It can learn any ratio, so it is the direct empirical competitor to "mechanism finds the right ratio."

Two ablations isolate the key ingredient. The first replaces the ODE with a generic latent-factor generative model fitted to the same wild-type data and then reruns the identical Bayes-AUC pipeline. This tests whether mechanistic structure matters or any fitted simulator would do. The second keeps the ODE but collapses the nuisance prior to a point mass. If the explanation is right, this ablation should rank panels by raw sensitivity and lose most of its correlation with empirical AUC.

Why not just…?

Why not just run LASSO or recursive feature elimination on the labeled data and take whatever panel wins? With enough labeled samples this will probably find panels as good as ours, and we do not claim otherwise. The case for the mechanistic criterion is narrower: it needs no labels, so it works when labeled samples are tiny (15 cell lines here), and it should be more stable under shift because it selects on structure rather than cohort-specific correlation. The 87% CTD-ILD composition and the 0.885 to 0.849 external drop in the ILD study show that empirical selection is cohort-bound. If nested LASSO matches our panels both in-distribution and under the amplification-level shift, the practical claim fails. What would remain is only the explanatory claim that identifiability accounts for which panels work, and we would report that.

Why not just compute ratios, as clinicians already do? For simple sinks, a ratio is what the mechanism recommends. The amyloid model should recover Aβ42/Aβ40 rather than improve on it. The ratio heuristic has no rule for which denominator to use, when two markers are needed, or when timing matters. The PI3K decoupling case is a test of this. If the nuisance is pathway expression and the disease effect appears as a change in the relationship between p-AKT and p-HER3 over time, no single static ratio may suffice. The log-contrast baseline will show directly whether a learned ratio is enough.

Why not just fit longitudinal mixed-effects curves to the time courses? Phenomenological slopes describe trajectories but do not separate a decoupling rate from an expression nuisance. They also cannot be queried about panels that were never measured, which is what panel design requires. The latent-factor ablation is the fair version of this alternative: a fitted generative model without mechanism. If it predicts panel quality as well as the ODE does, the mechanism is not earning its cost, and we would adopt the simpler model.

Decisive Checks

The program succeeds if three things hold. First, the Spearman correlation between simulated Bayes AUC and held-out empirical AUC across candidate RPPA panels is clearly positive, with a bootstrap interval over lines that excludes zero. Second, the point-mass-nuisance ablation and the latent-factor ablation both show markedly weaker correlation. Third, the panel the criterion picks without labels is not beaten by nested label-driven selection at the same panel size, and it degrades less under the amplification-level split. The amyloid model must reproduce the published ordering of Aβ42 alone, then Aβ42/Aβ40 (0.8343), with p-tau181 (0.8768) adding an independent axis, using priors that were not tuned to those numbers.

Failing the first check would mean identifiability does not explain diagnostic sufficiency in this system, and that is a publishable negative result. Passing the first check but failing the third would mean the idea explains panel quality without being useful for design, and the note's claim would shrink accordingly.